AI Disclosure

1. AI in AfterLight

AfterLight uses artificial intelligence as a core part of its reflection experience. Ben, the AI reflection companion, is powered by OpenAI (GPT-4o-mini). Ben generates reflective responses based on the memories you share. Ben is not a real person, does not simulate consciousness, and does not claim to have feelings or experiences.

2. What AI Does

The AI in AfterLight performs the following functions:

• Generates reflective responses to your messages based on shared memories • Helps organize and describe memories through structured conversation • Detects themes and patterns across your stored memories • Creates semantic embeddings (mathematical representations) of memory descriptions for similarity search • Proposes memory titles, groupings, and temporal sequences for your confirmation

3. What AI Does NOT Do

The AI in AfterLight does not:

• Simulate consciousness or sentience • Claim to be or represent a real person • Provide therapy, counseling, or medical advice • Diagnose or interpret your emotional state • Remember previous conversations (each session starts fresh) • Make decisions on your behalf • Learn or update its behavior based on your interactions (no fine-tuning) • Generate content involving real people without explicit user-provided context

4. How AI Uses Your Data

When you interact with Ben, your messages and relevant stored memories are sent to OpenAI's API for response generation. Under OpenAI's API data usage policy, this data is not used to train or improve their models. Embeddings (mathematical vectors) are generated from memory descriptions and stored in our database for semantic search — these are not human-readable.

Your data is transmitted to OpenAI servers in the United States for processing. All persistent data remains stored in the EU (Google Cloud, Belgium).

5. EU AI Act Classification

Under the European Union AI Act, AfterLight is classified as a limited-risk AI system. This means it is subject to transparency obligations under Article 50, but is not classified as high-risk or prohibited.

AfterLight is not high-risk because it does not:

• Make legal, employment, or credit decisions • Perform biometric identification • Operate critical infrastructure • Profile individuals for risk assessment

AfterLight is not prohibited because it does not use subliminal manipulation, exploit vulnerabilities, perform social scoring, or use real-time remote biometric identification.

6. Transparency Measures

We are committed to transparency about AI use in AfterLight:

• All AI responses are labeled "Ben (AI reflection)" in the chat interface • The chat header states: "Ben is an AI companion. Responses are generated, not personal." • A first-use acknowledgment screen informs new users about AI before their first conversation • This dedicated AI Disclosure page provides comprehensive information • Ben never claims to be a real person or to have consciousness

7. Ethical Guardrails

Ben operates within strict ethical boundaries:

• Ben never claims to be the person the user is reflecting on • Ben never speaks in the first person as that person • Ben never gives medical, psychological, or life advice • Ben never encourages emotional dependency • Ben never speculates about what a loved one would say or want • Ben never generates sexual, violent, or inappropriate content

These guardrails are enforced through Ben's system prompt and are regularly tested through automated persona simulations.

8. Models & Attribution

AfterLight uses the following AI models and research assets:

• OpenAI GPT-4o-mini — text generation for Ben’s reflective responses • OpenAI text-embedding-3-small — semantic memory embeddings • Google Cloud Text-to-Speech (Neural2) — voice narration • LivePortrait — neural micro-animation (photo-based motion synthesis) • DECA (Detailed Expression Capture and Animation) — 3D face reconstruction from a single photo • FLAME (Faces Learned with an Articulated Model and Expressions) — parametric 3D head model used by DECA

FLAME 2023 Open is used under the CC-BY-4.0 license. Attribution: "FLAME: Learning a model of facial shape and expression from 4D scans" by Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. © Max Planck Institute for Intelligent Systems.

DECA is used under the license terms of its original authors at the Max Planck Institute for Intelligent Systems.

9. Human Oversight

The Umbrella Research team monitors AfterLight's AI behavior through automated guardrail testing, persona-based stress tests, and ongoing review. If you encounter any AI response that seems inappropriate or concerning, please contact us at ethics@umbrella-research.org.